mac-use-mcp

Instructions

Implementation Reference

• src/tools/screenshot.ts:130-193 (handler)

  The handleScreenshot function executes the tool logic, including input validation, capturing the screen via `captureScreen`, and formatting the resulting image/metadata for the MCP response.

  async function handleScreenshot(
    args: Record<string, unknown>,
  ): Promise<CallToolResult> {
    const parsed = ScreenshotInputSchema.parse(args);
  
    try {
      const result = await captureScreen({
        mode: parsed.mode,
        region:
          parsed.mode === "region"
            ? { x: parsed.x!, y: parsed.y!, w: parsed.width!, h: parsed.height! }
            : undefined,
        windowTitle: parsed.mode === "window" ? parsed.window_title : undefined,
        maxDimension: parsed.max_dimension,
        format: parsed.format,
        ruler: parsed.ruler,
      });
  
      const mimeType = parsed.format === "jpeg" ? "image/jpeg" : "image/png";
  
      // Build coordinate mapping hint for agents
      const isIdentity =
        result.scaleX === 1 &&
        result.scaleY === 1 &&
        result.originX === 0 &&
        result.originY === 0;
  
      const coordinateHint = isIdentity
        ? "Coordinate mapping: screen = image pixel (1:1, no conversion needed)"
        : `Coordinate mapping: screen_x = ${result.originX} + image_x * ${result.scaleX}, screen_y = ${result.originY} + image_y * ${result.scaleY}`;
  
      return {
        content: [
          {
            type: "image" as const,
            data: result.base64,
            mimeType,
          },
          {
            type: "text" as const,
            text: `Image: ${result.width}x${result.height}\n${coordinateHint}`,
          },
        ],
      };
    } catch (error: unknown) {
      const message = error instanceof Error ? error.message : String(error);
  
      // Detect permission-related failures and include setup instructions
      const isPermissionError =
        /permission/i.test(message) ||
        /not permitted/i.test(message) ||
        /screen recording/i.test(message) ||
        /cannot be opened/i.test(message);
  
      const text = isPermissionError
        ? `Screenshot failed: ${message}\n\n${PERMISSION_INSTRUCTIONS}`
        : `Screenshot failed: ${message}`;
  
      return {
        content: [{ type: "text" as const, text }],
        isError: true,
      };
    }
  }

• src/tools/screenshot.ts:88-110 (schema)

  The ScreenshotInputSchema uses Zod for runtime input validation and includes cross-field refinements to ensure mandatory parameters for different screenshot modes (e.g., 'region' vs 'window').

  /** Full runtime validation schema with cross-field refinements. */
  const ScreenshotInputSchema = ScreenshotBaseSchema.refine(
    (data) => {
      if (data.mode === "region") {
        return (
          data.x !== undefined &&
          data.y !== undefined &&
          data.width !== undefined &&
          data.height !== undefined
        );
      }
      return true;
    },
    { message: 'x, y, width, and height are required when mode is "region"' },
  ).refine(
    (data) => {
      if (data.mode === "window") {
        return data.window_title !== undefined;
      }
      return true;
    },
    { message: 'window_title is required when mode is "window"' },
  );

• src/tools/screenshot.ts:114-125 (registration)

  The screenshot tool definition exported as part of screenshotToolDefinitions, used to register the tool in the MCP server.

  export const screenshotToolDefinitions: Tool[] = [
    {
      name: "screenshot",
      description:
        "Capture a screenshot of the macOS screen. Supports full screen, a rectangular region, or a specific window by title. Returns a base64-encoded image with dimension metadata. Do not narrate visual observations or coordinate calculations. Brief task progress updates are acceptable.",
      inputSchema: zodToToolInputSchema(ScreenshotBaseSchema),
      annotations: {
        readOnlyHint: true,
        destructiveHint: false,
      },
    },
  ];

• src/tools/screenshot.ts:198-203 (registration)

  The screenshotToolHandlers object maps the 'screenshot' tool name to its handler function, wrapping it in an enqueue function to manage execution flow.

  export const screenshotToolHandlers: Record<
    string,
    (args: Record<string, unknown>) => Promise<CallToolResult>
  > = {
    screenshot: (args) => enqueue(() => handleScreenshot(args)),
  };